NAND bad blocks
NAND memory apparently gets shipped with blocks that are already bad. The vendor just marks those blocks as bad, thus resulting in higher yield and lower per-unit cost. Thus, the flash will in the end contain three kinds of blocks:
- Factory-Default bad blocks
- Worn-out bad blocks
- Good blocks
The only block that is guaranteed to be good, is the first block (first 16kBytes).
We are also guaranteed that a minimum of 4016 blocks (out of the total 4096) are good. This means up to 80 blocks (1.3MBytes) can be dead, resulting in a total guaranteed amount of working NAND storage of 65798144 bytes.
The solution is split into various pieces, one at each level.
The boot loader itself contains a small first-stage boot loader for the S3C2410 Steppingstone.
This code (mostly written in ARM assembly) was altered to detect and skip bad blocks. This means, the first stage bootloader can itself extend over bad blocks.
This also menas that the flashing routine needs to detect and skip bad blocks, resulting in a u-boot image that can have gaping holes ;) The existing "traditional" sjf2410-linux JTAG flashing program is not detecting bad blocks (Note: this might be changed through a compile option, see below)
The u-boot environment is traditionally stored at a fixed location within the NAND flash. This is not acceptable, since it could be a factory-set bad block.
The solution that was implemented for OpenMoko/Neo1973 was to put the in-flash address of the environment into the out-of-band (OOB) area of the first block (the one which is guaranteed to be good). Since the environment address is unlikely to change often, the 1000 erase cycles guaranteed for the first block are good enough.
The exact location is byte 8..15 of the 16byte OOB area, starting with the four ASCII bytes ENVO, followed by the little-endian 32bit unsigned integer of the NAND address where the environment is located.
The u-boot "dynenv get" command can be used to read out a pre-programmed Environment offset from NAND, and the "dynenv set" can be used to write the offset (if the last eight bytes of OOB area are erased (0xff)).
Since those up to 80 factory-bad blocks can be located about anywhere in NAND, we have to accomodate for this worst case. However, we cannot just make every partition 1.3MB larger than it needs to be, since this would waste a lot of otherwise good flash.
The only solution to this (that Harald could think of) is to dynamically calculate a partition table for each device. Every NAND flash has different factory-bad blocks at different locations, thus the partition table on every NAND flash will look different.
So as an example, lets' assume we have a 0x30000 (196k) bytes sized partition for u-boot, starting ad address 0 in NAND. If there were no bad blocks, it would extend from 0x00000 to 0x30000. From 0x30000 to 0x230000 (2MB) we have the kernel partition.
Let's now assume that blocks 0x20000 and 0x28000 (each 0x4000 in size) are marked as factory-bad. Thus, in order to have 0x30000 bytes of usable storage, the uboot partition actually extends from 0x00000 to 0x38000. This shifts the start address of the kernel partition to 0x38000.
If the kernel partition contains more bad blocks, the start address of the rootfs partition (following the e kernel partition) is further shifted down to the end.
Those calculations have been implemented as u-boot "dynpart" command. Once you issue "dynpart", the partition configuration is put in the "mtdparts" environment variable. If you "saveenv" the environment, it is saved into the non-volatile environment partition.
Bad Block Table (BBT)
Since the usual bad block marker in the OOB area does not allow us to distinguish between factory-bad and worn-out-bad blocks, we need to store this information elsewhere. This places is called bad-block table (BBT), and it is placed as a bitmap into the last two working/good blocks at the end of NAND. To increase security, a backup of those two blocks is kept in the next two working/good blocks just before.
The BBT location itself is identified by special markers (BBT0/BBT1) in the OOB area of the first page of the respective erase blocks.
The BBT consists of two bits per block, which distinguish the three conditions (factory-bad/worn-out/good).
Both u-boot and Linux implement the same BBT layout and thus interoperate quite well.
The BBT is created once a BBT-implementing u-boot is started for the first time. The BBT scanning code assumes that the NAND is completely erased, i.e. only contains 0xff as content. Any block that contains bytes != 0xff in the OOB is marked as "factory bad" block.
The kernel is contained in its own partiton QT2410#NAND. We have to flash it using the
command, and read it later again via
command. Those two variants (as opposed to their non-".e"-postfixed versions) simply skip bad blocks
The root filesystem uses JFFS2, which is already bad block tolerant. It, too, has to be written using the
command in u-boot
sjf2410 (during development)
The sjf2410-linux tool has a compile-time option to check (and skip) bad blocks. If we use this for flashing u-boot, we will preserve the bad block info, once u-boot steppingstone code has been enhanced to skip bad blocks.
I have no idea how flashing during the production process is supposed to happen.
Tentative new Flash memory layout
start size name 0x00000000 0x00030000 196KB u-boot 0x00030000 0x00004000 16KB u-boot environment 0x00034000 0x00200000 2MB kernel 0x00234000 ... jffs2